Poster B82, Sunday, March 25, 8:00-10:00 am, Exhibit Hall C
Local Heterogeneity Regression Analysis: A Novel Measure of Representational Sparseness in Reading
Jeremy Purcell1, Brenda Rapp1; 1Johns Hopkins University
Orthographic representations become sparser after learning (i.e. strong activation in a relatively small set of neurons), and that differences in the mean neural response does not capture this relative sparseness (Glezer et al., 2009). Instead more advanced measures that quantify the local neural heterogeneity are required. Here we introduce a novel Local Heterogeneity Regression (Hreg) Analysis that quantifies the relative heterogeneity of the local neural responses. This approach relies on the premise that well-learned sparse representations will have heterogeneous local responses across voxels. We apply this approach to block design fMRI reading data (N=30) which included well-learned words (i.e. high frequency-HFW) and relatively less well-learned words (i.e. low frequency-LFW and pseudowords-PW). Local-Hreg is a search-light analysis where, for each search-light, a general psychophysiological interaction analysis (McLaren et al., 2012) is performed using the center voxel to make pair-wise predictions of each surrounding voxel. The Local-Hreg value is the median condition-specific pairwise interactions within a searchlight. Lower average condition-specific cross-voxel interactions indicate higher local heterogeneity. Results reveal a significant left vOTC cluster with higher heterogeneity for the well-learned words (HFW) relative to the less well-learned stimuli (LFW or PW). We compare Local-Hreg to another measure of heterogeneity (heterogeneity correlation-Hcorr), and confirm that it is a more robust method for detecting local heterogeneity. Here we introduce a novel approach for examining the relative sparseness of orthographic representations. We argue that it can be used as a general analysis tool for probing neural dynamics of representation and learning in various cognitive domains.
Topic Area: METHODS: Neuroimaging